Last updated on May 3rd, 2026 at 04:37 pm
By Jen Evans, founder of Pattern Pulse AI and B2BNN, co-founder of Tech Reset Canada
For good or ill, this is what we’ve got. No matter how the government and the minister of AI try to sell it, no matter how many NFPs or post-secondary-led proposals are submitted to SCIP, nothing is going to change an underlying fact: Canada’s AI strategy is not only being deployed well ahead of procurement guidelines, it is built on a triumvirate of some very shaky, even rotted underpinnings.
The trio: Cohere, CoreWeave, and Palantir. Two of them are American. The Canadian one occupies the smallest, most dependent role in the architecture.
The model layer: Cohere. Founded in Toronto in 2019 by Aidan Gomez (a co-author of the original “Attention Is All You Need” paper that gave us the transformer architecture), Ivan Zhang, and Nick Frosst (an early hire at Google’s Toronto AI lab). Cohere has raised roughly $1.5 billion at a $7 billion valuation, with investors including Nvidia, AMD Ventures, Salesforce Ventures, PSP Investments, BDC, HOOPP, Radical Ventures, and Inovia Capital. By the end of 2025 it had reached roughly $240 million in annual recurring revenue, with 85% flowing from private deployments to large enterprise customers: Oracle, Dell, RBC, Bell, Fujitsu, LG CNS, SAP. The flagship products are the Command family of generative models, the Embed and Rerank retrieval models, and an agentic AI platform called North.
Cohere does pretrain its own foundation models. The Command family is autoregressive transformer LLMs pretrained from scratch by Cohere and aligned through supervised fine-tuning and preference training. These are not assemblies of open source weights. What they are is something narrower than the word “frontier” implies. Command R+ is roughly 104 billion parameters. The 2026 frontier sits at GPT-5, Claude Opus 4.7, and Gemini, none of which Cohere contends with on raw capability. The 70% gross margins come from a deliberate decision to skip the speculative compute spend that frontier training requires, scaling in proportion to enterprise demand instead. The company has no consumer product. There is no Cohere chatbot a Canadian citizen can sign up for. The product is enterprise software, sold to regulated industries on multi-year private deployment contracts.
The contrast with Mistral is sharp. Mistral was founded in Paris in 2023 by Arthur Mensch (ex-DeepMind), Guillaume Lample, and Timothée Lacroix (both ex-Meta). It has raised roughly $2.7 billion at a $13.7 billion valuation, with ASML now the largest shareholder at 11%. Mistral ships open-weight models under Apache 2.0 (Mistral 7B, Mixtral, Mistral Large 3, the Ministral and Magistral families), operates a consumer chatbot called Le Chat that reached a million downloads in its first fourteen days, and is building sovereign datacenter capacity near Paris and in Sweden with $830 million in capex committed. Mistral Large 3 is a frontier-class mixture-of-experts model with 41 billion active and 675 billion total parameters, multimodal and multilingual, available for self-hosting on a single GPU. The stated mission is to offer a credible alternative to the closed-weights AI oligopoly.
The contrasts compound on every axis that matters for sovereignty. Mistral has open weights; Cohere does not (Cohere Labs releases some research weights for non-commercial use, but the commercial production models are proprietary). Mistral has a consumer product accessible to French and European citizens; Cohere has none. Mistral is building European compute capacity directly; Cohere depends on partner clouds and on the CoreWeave Canadian datacenter buildout. Mistral benchmarks against OpenAI and Anthropic; Cohere has deprioritized that competition in favour of enterprise margin.
That is the Cohere reality. It is a corporate model company built for corporate deployment work, not a frontier model lab built for public-sector governance work. Cohere has built real revenue and real partnerships. There is nothing structurally wrong with that as a business strategy. The problem arrives when “Canada’s leading AI company” is asked to carry the Canadian sovereignty position and the public-sector governance load that goes with it. The strategy that makes Cohere capital-efficient (skip frontier scale, lean on partner cloud, sell into enterprise) is the same strategy that makes Cohere structurally insufficient as a sovereignty backstop. Canada’s domestic model champion is, by design, a B2B vendor with no public-facing model, no Canadian compute of its own, and no public-sector governance posture.
The compute layer: CoreWeave. The US-listed specialized GPU cloud provider, originally a crypto miner, owns and operates datacenters racked with Nvidia H100, H200, and B200 GPUs. Microsoft is its largest customer. OpenAI’s compute flows through CoreWeave via Microsoft. CoreWeave is Cohere’s datacenter partner on the multibillion-dollar Canadian buildout that the federal C$240M commitment under the Canadian Sovereign AI Compute Strategy anchors.
CoreWeave’s growth is debt-fueled, and the debt is below investment grade. S&P rates the company B+. As of late 2024, CoreWeave carried roughly $8 billion in debt at interest rates between 11% and 15%, with $7.5 billion in debt and interest payments due by the end of 2026. In April 2026 alone, the company raised roughly $5.75 billion across three high-yield debt sales in a single week: $1.75 billion of 9.75% senior unsecured notes due 2031 plus $3 billion of convertible debt on April 9, and an additional $1 billion tap of the same senior notes on April 16. The Canadian sovereign compute footprint sits on a counterparty whose capital structure is being held up by serial issuance into a frothy AI debt market.
The analytics layer: Palantir. The US data and software company, founded in 2003 with In-Q-Tel funding, sells two flagship platforms (Gotham for defense and intelligence, Foundry for civilian government and commercial use) plus an Artificial Intelligence Platform that wraps third-party LLMs inside Palantir’s Ontology framework. Palantir does not train foundation models. The Canadian federal footprint runs through a $14.4 million Department of National Defence data management contract from March 2020, surfaced in Parliament in late 2025, alongside ongoing analytics procurement in other departments.
On April 18, 2026, Palantir posted a 22-point manifesto on X drawn from CEO Alex Karp and corporate affairs head Nicholas Zamiska’s book The Technological Republic. The document has been widely characterized as technofascist by political scientists and major outlets, including The Nation, MS Now, Al Jazeera, Common Dreams, TechPolicy.Press, and Byline Times. Dutch political scientist Cas Mudde, who studies authoritarianism, called it “one of the scariest things I have seen in a while” and recommended European countries end any reliance on Palantir for security. University of Michigan political scientist Donald Moynihan concluded that the manifesto’s vision is “that of a US government and its tech allies as dominant players, unconstrained by accountability.” The document calls for an “engineering elite” with an “affirmative obligation to participate in the defense of the nation,” disparages soft power and democratic deliberation as inefficient, and frames AI weapons as both inevitable and morally obligatory. This is the political position published by the company holding active Canadian federal data and analytics contracts.
The triumvirate is the operating reality of Canadian sovereign AI. Cohere supplies the model weights the Canadian narrative claims as its own. CoreWeave supplies the physical compute that runs those weights. Palantir supplies the application layer that ingests the data and operationalizes the outputs inside government workflows. Two of those three companies are American. The Canadian one is tethered to the American compute provider for its largest national infrastructure commitment, has no public-facing product, and has chosen not to compete at the frontier.
Stack the three counterparties side by side and the sovereignty argument gets harder, not easier. A B2B vendor that has explicitly opted out of the frontier and out of public-facing deployment. A debt-fueled GPU host running a serial junk-bond program to finance expansion, with $7.5 billion in debt and interest payments due before the end of this year. A US data company whose CEO has just published a 22-point manifesto that scholars of authoritarianism are calling technofascist, while holding active Canadian federal contracts. The procurement question for the 2028 Supply Arrangement renewal is whether Canada is willing to name what it is buying, or whether the word “sovereign” continues to do the work of papering over it.